基于Swin Transformer的岩石岩性智能识别研究

Research on rock lithology intelligent identification based on Swin Transformer

常规卷积神经网络在识别纹理多变的岩石图像时,由于感受野和局部处理方式的局限性,识别精度不高,为解决上述问题,在复杂情况下准确识别岩石岩性,提高地质调查的效率,文中提出一种基于改进Swin Transformer的岩石识别方法。该方法增加了空间局部感知模块,并结合Transformer的自注意力结构来增强对局部相关性的提取。为增强泛化,模型中添加了Dropout层,减少对单神经元的依赖。为进一步提高网络的泛化能力,采用AugMix算法对岩石图像进行数据增强,并结合迁移学习技术对网络进行预训练从而优化网络参数。实验结果表明,该方法的识别准确率为96.4%,高于ResNet50、GoogLeNet、VGG16网络。

Due to the limitations of their receptive fields and the method of local processing,conventional convolutional neu-ral networks(CNNs)struggle with high recognition accuracy when identifying rock images with varied textures.In order to accu-rately identify rock lithology in complex scenarios and thereby improve the efficiency of geological surveys,a rock identification method based on improved Swin Transformer is proposed.A spatial local perception module is incorporated into this method.The Transformer's self-attention mechanism is combined to enhance the extraction of local correlations.To enhance generalization,a Dropout layer is added to the model,which reduces the dependence on individual neurons.To further enhance the generalization capability of the model,the AugMix algorithm is employed for rock image data augmentation.In combination with transfer learning technology,the network is pre-trained to optimize its parameters.Experimental results show that the accuracy of the proposed method reaches 96.4%,which outperforms that of networks like ResNet50,GoogLeNet and VGG16.